Legal AI That
Stays Confidential.

Every weight in the BitNet b1.58 backbone is ternary: −1, 0, or +1. A weight of +1 is a copy, 0 is a skip, −1 is a negation. The CPU's multiply unit idles entirely; matrix multiplication becomes POPCNT (population count of set bits) and XOR — operations that draw a fraction of the energy of a standard FMA pipeline. This is why Vexis hits a 1.5–3 W active power envelope on mid-range silicon.

Two compression layers. NanoQuant decomposes weight matrices as W ≈ α·(B₁ · B₂ᵀ) — binary factors with a learned scale vector — bringing 10B ternary weights from ~1.975 GB down to ~820 MB at 0.082 bytes per parameter. TurboQuant then applies 3-bit per-head quantization to the KV cache, and KeyDiff evicts low-relevance entries token-by-token. Total budget: ~1.17 GB with 130 MB headroom.

Mid-range Android targets are Litigation (5–7 tok/s), Contract Law (7–10 tok/s), and Criminal Defence. Flagship devices like Corporate Law reach 12–18 tok/s with 8K context. ExecuTorch dispatches to NPU via QNN, ANE, or APU delegates; SIMD kernels cover ARM NEON, x86 AVX2, and RISC-V Vector. Vulkan compute shaders accelerate Family Law GPUs by 2–3×.

Self-Assessment & Retry. A distilled 300M verifier classifies each response into one of five domains — Code, Math, Legal, Financial, or General — then generates 3–5 rubric questions and grades the candidate output against them. Failed checks trigger a retry at temperature 0.2 (max 3 attempts). It costs 120–200 ms per response but lifts TruthfulQA by +3.8% over the FP16 baseline. Trustworthiness over raw speed is a deliberate product decision.

Three sovereignty levels. L0 (Air-Gap) ships zero network traffic — TC-38 verified via Wireshark, suitable for healthcare and enterprise. L1 (Signal-Only) transmits binary quality signals only, no text. L2 (Opt-in Snippet) sends NER PII-scrubbed snippets through a manual review dashboard. Default consumer mode is L1. Memory is local to your device across three tiers: profile.txt, memory.txt, and daily_log.txt.

BitNet b1.58 alone holds ~96.2% of FP16 accuracy. Kolmogorov-Arnold Network layers — replacing the transformer's MLP feed-forward blocks — close most of the remaining gap by placing learnable univariate spline activations on edges rather than fixed activations on nodes. The combination hits ~98.5% of FP16 accuracy. We use the KAT variant with grouped KANs and rational bases for NPU efficiency, with Wav-KAN (wavelet) as a 1.2–1.8× speed alternative.

Yes. Sub-layer LoRA rank-8 adapters with Straight-Through Estimator gradients let you fine-tune the ternary backbone on-device or server-side. Each transformer block keeps a frozen ternary base plus a trainable FP16 adapter — ~120 MB total for the full 10B model. Training reaches 98% of full fine-tune accuracy at rank-8, completing in under 30 minutes on-device.

The public API exposes VexisEngine.init(), chat(), setMemoryTier(), setSovereigntyLevel(), and registerAdapter(). Internal primitives — inference dispatch, the SAR verifier, KeyDiff eviction — are encapsulated. Ships with developer documentation and integration guides as part of the Phase 6 launch package.